fix(decisioning): TaskHandoff registry.fail/complete echo request context

[bokelley]

Closes #563 (follow-up to #560).

The gap

PR #560 fixed context-echo on the synchronous AdcpError raise path across MCP and A2A. Background tasks (TaskHandoff) route through a separate path that #560 didn't touch: `registry.complete(task_id, result)` on success and `registry.fail(task_id, error.to_wire())` on AdcpError (or wrapped INTERNAL_ERROR). Neither echoed the request's `context` extension, so a buyer polling `tasks/get` on a deferred task lost their correlation IDs across the handoff boundary.

The fix

`_project_handoff(...)` now accepts `request_params: BaseModel | None`. Both terminal-state helpers — `_fail` and the success branch that calls `registry.complete` — apply `inject_context(_to_request_dict(request_params), wire_dict)` before passing to the registry.

The echo is now symmetric across the four terminal-state paths:

Path	Site	Status
Sync MCP / A2A success	`mcp_tools.py:2030` / `a2a_server._send_result`	already echoed (#560)
Sync AdcpError raise	`serve.py` via `build_mcp_error_result`	already echoed (#560)
Bg task success	`registry.complete` in `_project_handoff`	this PR
Bg task failure (incl. wrapped INTERNAL_ERROR)	`registry.fail` via `_fail` in `_project_handoff`	this PR

New helper

`_to_request_dict(params)` coerces a Pydantic request model (or dict) to a plain dict for `inject_context`. Empty-dict fall-through on coercion failure — the echo is a feature, not a load-bearing invariant; if Pydantic fails to dump, the wire envelope still goes out without crashing.

Tests

4 new in `test_decisioning_dispatch.py`:

• success-path echo: `registry.complete` carries `context` from the request
• AdcpError-path echo: `registry.fail` carries `context` alongside `adcp_error`
• wrapped-INTERNAL_ERROR echo: raw `RuntimeError` → wrap → `registry.fail` envelope still echoes
• no-request-params no-op: when called from test fixtures without `request_params=`, no synthesised context

42 dispatch tests green; ruff clean.

Test plan

• `pytest tests/test_decisioning_dispatch.py` — 42 passed (38 existing + 4 new)
• `ruff check` — clean

🤖 Generated with Claude Code

[bokelley]

Updated based on expert review (code-reviewer + security-reviewer + ad-tech-protocol-expert):

Spec-correctness fix (protocol-expert): Original approach injected context inside the result/error dict. Per schemas/cache/core/tasks_get_response.json, context is a top-level sibling of result/error/history. Refactored so:

• TaskRecord gains a request_context field; to_dict() surfaces it at the top level under context.
• TaskRegistry.issue() accepts request_context=None kwarg (additive, default backwards-compatible).
• _project_handoff extracts context from request_params and passes to issue() once at issue-time. Terminal-state helpers (_fail, registry.complete) no longer mutate the wire dict.

Workflow_handoff coverage (reviewer #2): The same context-echo gap existed in _project_workflow_handoff — that path was missed in v1 of this PR. Now also threads request_params and calls issue(request_context=...).

Observability fix (security-reviewer): _to_request_dict's silent except Exception → {} is now _extract_request_context and logs a warning when model_dump fails. Buyer correlation IDs being lost are now detectable in operator logs.

Tests updated to assert top-level rec.context (not nested) and 'context' not in rec.result / rec.error to pin the spec invariant.

Open question I deferred (security-reviewer L1): credential-shaped keys in the buyer's request.context field are not currently gated by _validate_ctx_metadata_credentials. Buyer-leaks-to-themselves contract is preserved (registry is account-scoped via expected_account_id), but if tasks/get ever gains a cross-account fan-out (admin / governance read), buyer's context becomes an exfil channel. Left as a follow-up to track separately — not new in this PR.

129 tests across registry/idempotency/dispatch green; lint clean.

[bokelley]

Thanks for the detailed update. The spec-correctness refactor (context as top-level sibling rather than injected inside result/error), the _project_workflow_handoff coverage gap fix, and the observability improvement all look like solid improvements over the original approach. The deferred credential-leakage follow-up is a reasonable call — the account-scoping invariant holds today, and tracking it separately keeps this PR reviewable.

Ready for human review when you are.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01WyXp7wGvN7oX9VVcwxt9iR

---

Generated by Claude Code

[bokelley]

Addressed the P3 nit: TaskRegistry.issue() now declares **_extra: Any (both Protocol and InMemoryTaskRegistry impl). Custom registry impls in the wild that haven't yet adopted request_context= won't TypeError when the framework introduces new optional kwargs — they silently absorb unknowns and the in-memory impl logs at DEBUG so adopters notice they've fallen behind.

Documented the forward-compat contract in the Protocol docstring: custom impls MUST include **_extra: Any on their issue() signature.

Tests still green. Ready for admin-merge.

[bokelley]

The **_extra: Any forward-compat contract is a clean solution — custom registry impls that lag behind the framework's kwarg additions won't TypeError, and the DEBUG log gives operators a signal without being noisy. Looks good from a triage perspective; nothing new to flag.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_013aJ8fiv2FgJedEFoBEuHrV

---

Generated by Claude Code